Master Art of Translation | The Translation Unit — How to Decide What Must Move Together Across Languages

Translation units, units of translation, translation segmentation, sentence-by-sentence translation, phrase translation, CAT tool segments, translation memory units and AI translation chunks all point to one deceptively simple question: how much source text should be treated as one meaning-bearing object before it is rebuilt in another language? People searching for how to translate phrases instead of words, how to split sentences for translation, how translation memory segments work, how to preserve context in machine translation, or why a word-for-word translation sounds wrong are usually meeting the same underlying problem. The source has been cut at the wrong boundary.

This guide explains how to choose a translation unit without confusing a linguistic unit with a software segment. It shows when a word is enough, when a phrase must stay intact, when a clause carries the real relation, when two sentences must be understood together, and when the correct unit is a whole paragraph, interface state, subtitle exchange, legal provision, technical procedure or rhetorical movement. It also connects translation-unit decisions to translation memory, terminology, CAT tools, machine translation, AI translation, quality assurance, context, grammar, vocabulary, collocation, idioms, cohesion and reader purpose.

The practical principle is this: translate the smallest unit that contains enough evidence to preserve the required meaning and function, but never make the unit so small that dependencies are cut apart. That principle is useful for human translators, students, editors, localisation teams, AI users and anyone translating between languages that package meaning differently. Good translation is not merely choosing good target words. It is deciding which source elements belong to one decision.

What this architecture node owns

This article owns the architecture-level question of translation units and segmentation. The earlier Master Art of Translation source-analysis node owns how to read the source before translating. The equivalence node owns what it means for languages to correspond without matching one-to-one. The context-stack node owns the layers of co-text, document, situation, culture and interface that constrain meaning. This node asks a different question: once those layers are understood, what should be translated together?

It does not replace eduKateSG vocabulary teaching, English grammar teaching, specialist terminology pages, rapid chunking techniques, translation-memory tutorials, localisation pages or later genre branches. It uses those systems when they help solve the unit problem. Vocabulary can explain the semantic range of a word. How English Works can explain the syntax of a clause. The translation unit decides whether that word or clause can safely be moved independently at all.

The hidden problem: boundaries create meaning

Every translation begins with boundaries, even when the translator never names them. A reader sees spaces between words, punctuation between sentences, line breaks between paragraphs, cells in a spreadsheet, strings in a localisation file, subtitle cues on a timeline and segments in a CAT tool. These visible boundaries are useful, but they are not guaranteed to be meaning boundaries.

Consider the English expression “kick the bucket.” The spaces say three words. The meaning may be one idiomatic unit. Translating kick, the and bucket independently can destroy the intended sense. Now consider “The committee approved the plan. However, implementation was delayed.” The period divides two sentences, but the second sentence carries a contrast relation that cannot be understood properly without the first. A software tool may show two segments. The translator still needs one discourse model.

The hidden problem is therefore not segmentation itself. Segmentation is necessary for attention, workflow, storage, reuse and editing. The problem is mistaking the segment boundary for the edge of meaning. A segment is where a tool or document has cut. A translation unit is the span a translator must treat as one coordinated decision.

What is a translation unit?

In this guide, a translation unit is the smallest source span that must be interpreted together so that its target rendering can preserve the relevant meaning, relation and function. The unit can be smaller than a sentence, equal to a sentence, or larger than a sentence. Its size changes by problem.

A multiword term such as “machine learning model” may behave as one unit even though it contains three words. A conditional clause may need to be paired with its result clause. A pronoun may require a previous sentence for reference. A joke may depend on setup delivered several lines earlier. A legal definition may control a term throughout a contract. A button labelled “Continue” may require the entire user-flow state to be understood before one target word can be chosen.

This makes the translation unit dynamic. The translator does not choose one permanent size for the whole profession. The translator chooses the unit that fits the dependency being solved.

Translation unit versus CAT-tool segment

Professional translation software often divides source files into segments, commonly sentences, titles or other units defined by segmentation rules. Phrase documentation describes source text being segmented and confirmed source-target pairs being stored in translation memory as translation units. The tool-level translation unit is therefore a stored source segment and its target counterpart, often with context metadata.

That operational definition is useful for memory and reuse, but a human reasoning unit can cross the displayed segment boundary. A translator may need to inspect the previous and next segment to resolve a pronoun, select a term, maintain parallel structure or understand a joke. Phrase also documents context metadata such as neighbouring segments or segment keys, and memoQ allows translators to join or split segments when segmentation does not match the material. These tools embody an important fact: segmentation is editable because document boundaries and translation decisions do not always coincide.

Never let the software’s row become your theory of meaning. The row is a workspace. Meaning may be narrower or wider.

The unit ladder

It helps to think of translation units as a ladder of possible scales. At the smallest end are morphemes and word parts. Then come words, compounds, fixed expressions, collocations, phrases, clauses, sentences, sentence pairs, paragraphs, sections, documents, interfaces, scenes and whole communicative events. The translator moves up or down the ladder until the relevant dependency fits inside the chosen window.

A prefix may matter when translating technical morphology. A word may be enough when a label has one stable equivalent in a well-defined domain. A phrase may be necessary for an idiom. A clause may be necessary for scope and condition. A sentence may carry a complete proposition. A pair of sentences may be needed for contrast. A paragraph may be required for reference or argument. A whole document may be required for terminology policy. A user flow may be required for an interface string. A scene may be required for subtitle dialogue. The right scale follows the problem, not the punctuation.

The dependency envelope

The most useful mechanism for choosing a translation unit is the dependency envelope. Imagine drawing a boundary around every source element that must be seen together to make one translation decision safely. The envelope expands until all necessary dependencies are inside it.

If the problem is a phrasal verb, the envelope may contain two or three words. If the problem is a relative pronoun, the envelope may contain a full clause. If the problem is who “she” refers to, the envelope may include several sentences. If the problem is whether “charge” means a fee, an accusation, an electrical property or battery replenishment, the envelope may need domain context from the section or document. If the problem is a one-word software label, the envelope may include the surrounding screen and action that follows the click.

A good translation unit is therefore not simply “a chunk.” It is a dependency envelope sized to the decision.

The minimum-sufficient-context rule

Use the smallest unit that is sufficient, not the smallest unit that is visible. This matters because very large units create their own problems. If every decision requires rereading an entire document, translation becomes slow and inconsistent. If every sentence is translated with the whole novel mentally active at once, attention becomes overloaded. Professional work needs manageable units.

The minimum-sufficient-context rule balances precision and cognitive load. Start small. If the meaning is stable, translate. If a dependency escapes the boundary, expand the unit. When the dependency is resolved, stop expanding. This creates a repeatable discipline: phrase, sentence, local co-text, paragraph, section, document, domain, world.

Signal 1: the unit is too small when the words change sense together

Idioms, phrasal verbs, multiword terms and fixed expressions are classic warnings. “Carry out” may mean perform. “Set up” may mean establish, configure, arrange or trap, depending on context. “Hot dog” is not usually translated as an adjective plus an animal. “Due process” is a legal concept, not merely two ordinary words. “Learning outcome” often functions as a domain term. If translating each word produces a plausible but strange result, test whether the words form a larger lexical unit.

The diagnostic question is: does the whole mean something that cannot be recovered by independently combining the most obvious dictionary senses? If yes, expand the unit.

Signal 2: the unit is too small when grammar determines scope

Negation, quantifiers, conditionals, comparatives and modifiers often require clause-level units. “Not all students passed” is not equivalent to “No students passed.” “Only registered users may enter” differs from “Registered users may only enter.” “If payment is not received within thirty days, access may be suspended” contains condition, time limit and modal force. Breaking these into isolated word decisions invites scope errors.

When meaning depends on who modifies what, what is negated, what is conditional or what is being compared, make the clause the unit.

Signal 3: the unit is too small when reference points outside it

Pronouns, demonstratives and elliptical expressions often depend on earlier language. “It was rejected” is not safely translatable without knowing what “it” refers to. “This caused the delay” requires the previous proposition. “The former” and “the latter” require a preceding pair. In languages that encode gender, number, animacy or honorific status differently, reference may force distinctions that the source did not make explicit.

If the current unit contains a pointer but not its target, expand until the target is included.

Signal 4: the unit is too small when a connective expresses a relation

Words such as however, therefore, although, because, instead, meanwhile, nevertheless, similarly and consequently derive their function from a relation between propositions. A translator who sees “however” as a vocabulary item without examining what is being contrasted may choose an unnatural or misleading connector. In some languages, the relation may be expressed through word order, particles, verb forms or no explicit connective at all.

For discourse connectors, the translation unit usually includes both sides of the relation.

Signal 5: the unit is too small when tone accumulates

Tone may not be stored in one word. A polite refusal can emerge across several sentences: appreciation, difficulty, alternative and closing. A character’s sarcasm may depend on a previous event. A research abstract’s caution may accumulate through repeated hedges. A brand voice may depend on short sentence rhythm across a paragraph. Translating one sentence in isolation can produce local naturalness and global tonal inconsistency.

If the emotional or interpersonal effect is cumulative, use a larger unit for review even if individual sentences can be drafted separately.

Signal 6: the unit is too small when terminology is defined elsewhere

A defined term can be one word at the point of use and a document-wide unit at the point of interpretation. “Provider,” “controller,” “consideration,” “material,” “cell,” “field,” “model” and “class” can have ordinary and specialist senses. A glossary, definition section, specification or curriculum may establish the intended concept. The translator must reach beyond the local sentence to recover that definition.

Terminology therefore has two unit sizes at once: the expression may be small, but its evidence base may be document-wide or domain-wide.

Signal 7: the unit is too small when layout supplies meaning

A table header, chart label, menu item, form field, button or warning icon may be impossible to translate safely from text alone. “Open,” “Close,” “Home,” “Back,” “Save,” “Apply,” “Free,” “Charge” and “Current” all change meaning across interfaces. The visible string may be one word, but the real unit includes the screen, control type, user state and resulting action.

For interface translation, the smallest visible string is often not the smallest sufficient unit.

Signal 8: the unit is too small when the reader’s task crosses the boundary

Instructions, legal provisions, recipes, experiments and procedures are action systems. A sentence may describe preparation and the next sentence a condition. A warning may apply to several following steps. A contractual definition may govern a later clause. If the reader must combine several segments to act correctly, the translator should review them as one functional unit.

When the unit is too large

Oversized units cause different failures. They can hide terminology inconsistency, make reuse difficult, overload attention, encourage paraphrase, blur alignment and reduce the usefulness of translation memory. A paragraph may contain several independent instructions that should remain individually testable. A long sentence may contain two interface strings that need separate localisation. A literary paragraph may contain a quotation requiring its own voice.

The solution is not always “translate more context.” It is “use enough context while keeping decisions inspectable.” The translator can understand a paragraph globally while drafting and reviewing at sentence or clause level. Context window and output unit do not have to be identical.

Understanding unit versus production unit

This distinction solves many apparent contradictions. The understanding unit is the span you need to interpret correctly. The production unit is the span you actually write or confirm at one time. You may need a whole paragraph to understand a pronoun chain, but produce one sentence at a time. You may need the entire screen to understand a button, but output one word. You may need a full contract to interpret a defined term, but translate one clause at a time.

When people argue about whether translators should work word by word, phrase by phrase or sentence by sentence, they often mix these two units. A translator can draft sentence by sentence while reasoning across paragraphs. Good workflows make that wider reasoning easy.

Storage unit versus reasoning unit

Translation memory adds a third distinction: the storage unit. A CAT tool may store a source segment and target segment as a translation unit for future matching. That storage choice matters because it shapes reuse. If the segment is too broad, future matches become rare. If it is too narrow, the stored translation may lose context and reappear incorrectly elsewhere.

Phrase documentation notes that translation-memory entries can carry metadata such as domain, client, file and context, and that context can be based on neighbouring segments or keys. This is a practical response to the weakness of isolated matching: the same source words can need different target renderings in different contexts.

The three-unit model

A robust workflow distinguishes three units: the understanding unit, the production unit and the storage unit. The understanding unit can be broad. The production unit should be manageable. The storage unit should be reusable without becoming dangerously context-free.

These units sometimes coincide. A clear standalone sentence can be understood, translated and stored as one segment. They sometimes diverge. A one-word button may require a whole screen to understand, one word to produce, and a keyed UI string with metadata to store. A legal term may require a whole definition section to understand, one term to produce and many source-target occurrences to manage through a termbase rather than ordinary translation memory.

The Unit Gate: a repeatable method

Before translating a difficult segment, run the Unit Gate.

Step 1: identify the visible segment. What did the document or software place in front of you?

Step 2: identify the decision. Are you choosing word sense, resolving grammar, preserving tone, applying terminology, maintaining reference, fitting layout or performing a reader action?

Step 3: find dependencies. Which words, clauses, sentences, definitions, visuals or real-world facts determine that decision?

Step 4: expand until the dependencies fit. Move from phrase to sentence to co-text to paragraph to section to document to interface or domain only as far as necessary.

Step 5: choose the production unit. Decide how much target text you can draft coherently at once.

Step 6: verify against the larger context. Make sure the local solution still fits terminology, register, cohesion and purpose.

Step 7: store or record the decision at the appropriate level. Use translation memory for reusable segments, a termbase for concepts, a style guide for recurring language policy, and a decision log for exceptional choices.

Failure mode: the dictionary-word unit

The simplest failure is treating every space-delimited word as a separate translation problem. This creates literal idioms, broken phrasal verbs, wrong compounds and unnatural collocations. The translator can know many words and still fail because the source packages meaning above the word level.

The repair is to test whether neighbouring words constrain sense. If they do, promote the unit from word to phrase or clause.

Failure mode: the punctuation-unit trap

A period is a useful segmentation mark, but not a guarantee of semantic independence. “He declined. Not because he disagreed, but because the timing was wrong.” The second sentence is structurally dependent on the first. Translating it without the first can create an incomplete or unnatural target. Headings and fragments often depend on surrounding material even more strongly.

The repair is to distinguish sentence boundary from dependency boundary.

Failure mode: the screen-row trap

CAT tools encourage focus by presenting one row at a time. This is productive until the row becomes a tunnel. Translators may confirm a segment before seeing a later definition, create inconsistent references across adjacent sentences, or accept a translation-memory match whose context differs.

The repair is disciplined context scanning: previous segment, next segment, paragraph, heading, document preview and relevant metadata. The row is a work surface, not a semantic prison.

Failure mode: the translation-memory match trap

A 100% string match does not automatically mean a 100% contextual match. The same sentence can appear under different headings, products, user roles or legal regimes. Translation-memory systems therefore use context and metadata to improve matching. Human translators should do the same conceptually.

Before accepting a high match, ask whether the source segment plays the same function, refers to the same entities, uses the same terminology and appears in the same locale or product context. Reuse is powerful because language repeats. It is dangerous for exactly the same reason.

Failure mode: segmentation that cuts a multiword term

File formats, line breaks, tags or punctuation can divide a multiword term. A product name may span styled text. A legal phrase may contain a cross-reference. A sentence may be split at an abbreviation. If the segmentation engine cuts inside a conceptual unit, the translator can misread the pieces.

Use document preview, tag awareness and join/split functions where permitted. The correct translation may require reuniting what the file structure separated.

Failure mode: joining too much

Joining segments can repair meaning, but indiscriminate joining reduces reuse and may make future editing harder. A whole paragraph stored as one unit is less likely to match another text. It also becomes difficult to isolate one changed sentence. Join when the content is genuinely interdependent or when the tool has split incorrectly, not merely because reading a long block feels easier.

Failure mode: splitting for convenience

Splitting a long sentence can improve attention, but the target may need to be reassembled with dependencies intact. If one half contains a condition and the other the result, or one half introduces a noun and the other a restrictive clause, treating them as independent can change meaning. Use temporary cognitive subunits without forgetting the larger grammatical unit.

Words as translation units

Sometimes a word really is enough. Stable technical labels, days of the week, common menu commands, chemical element names, basic numerals and well-defined glossary terms may have straightforward equivalents under a fixed brief. Even then, the apparent simplicity comes from established context. A word is safe as a unit when its sense, grammatical role, register and function are already determined.

The word unit is therefore a result of resolved context, not proof that context was unnecessary.

Morphemes and word parts as units

Some languages encode information inside words that another language distributes across several words. Prefixes, suffixes, inflections and compounds can carry tense, number, politeness, case, evidentiality, negation or derivation. A translator may need to analyse a single source word into multiple semantic components before producing a phrase or clause in the target.

This is one reason the visible word is not a universal linguistic unit. Languages package meaning differently. Translation can move from one word to many words or from many words to one without losing equivalence.

Compounds as units

English compounds such as “data protection officer,” “school attendance policy,” “battery charging indicator” and “student performance review” can hide relationships between nouns. Another language may require prepositions, adjective forms, possessives or clauses. Before translating, expand the compound mentally: officer responsible for data protection; policy governing school attendance; indicator showing battery charging; review of student performance.

The compound becomes a translation unit because its internal relationship must be solved before any target wording is chosen.

Collocations as units

Collocations are combinations that competent speakers prefer: make a decision, heavy rain, strong evidence, commit an offence, draw a conclusion, pose a risk. Another language may choose different verbs or adjectives. Translating the parts separately can create grammatical but unidiomatic text.

When one word predicts the natural partner, translate the collocation as a unit. This is where Vocabulary knowledge supports translation without taking over the translation owner’s job. Vocabulary can teach collocational range; translation decides which target combination fits the current source function.

Idioms as units

Idioms are obvious candidates for phrase-level translation, but the key question is whether the source is using the expression idiomatically, literally or both. “Break the ice” at a meeting differs from breaking ice on a frozen surface. “On the fence” may be figurative or physical. A literary text may exploit both senses.

The translation unit must include enough scene context to classify the phrase before a target equivalent is chosen. A phrase can be the production unit while a paragraph or scene is the understanding unit.

Phrasal verbs as units

English phrasal verbs frequently resist word-by-word translation. “Carry out,” “look up,” “turn down,” “work out,” “bring up,” “take over” and “set off” can each have multiple senses. The particle is not decorative. It participates in the lexical meaning.

Translate the verb-particle construction only after selecting its contextual sense. The target may use one verb, a phrase, a separable construction or a completely different metaphor.

Clauses as units

Clauses are often the smallest reliable units for logical meaning. Subject, predicate, objects, modifiers, tense, aspect, modality, negation and condition interact inside the clause. A source language may organise these relationships through word order while the target uses morphology or particles. The translator may change surface structure while preserving the clause’s semantic roles.

If a sentence contains several clauses, analyse each relation first, then decide whether the target should preserve, merge or split the source sentence.

Sentence as default working unit

The sentence is a practical default because it often contains a complete proposition and because translation tools commonly segment at sentence boundaries. It is large enough to preserve much syntax and small enough to review, align and reuse. But “default” is not “law.” The translator should always be ready to look outside the sentence or inside it more closely.

Sentence pairs as units

Contrast, explanation, cause, correction, question-answer structure and rhetorical setup often cross sentence boundaries. “The trial failed. Not because the method was wrong, but because recruitment stopped early.” The second sentence cannot be translated well without the first. “Did the update fix the problem? Not entirely.” The answer depends on the question.

When one sentence is structurally parasitic on another, treat them as one understanding unit.

Paragraph as unit

Paragraph-level translation becomes important for cohesion, topic development, reference chains, information flow and tone. A paragraph may introduce a concept, develop it through examples and end with a conclusion. Languages differ in how they repeat nouns, use pronouns, order old and new information and mark transitions. Sentence-by-sentence literalness can produce a paragraph that feels disjointed even when every sentence is individually correct.

Draft locally, review globally. The paragraph is often the minimum unit for natural discourse.

Section as unit

Sections control terminology and rhetorical purpose. A “Warnings” section has different language expectations from “Overview.” A “Methods” section differs from “Discussion.” A legal “Definitions” section establishes future usage. A product “Troubleshooting” section prioritises symptom-action structure. Understanding the section tells the translator how individual sentences function.

Document as unit

Some decisions are document-wide: terminology, names, spelling variety, address form, date format, defined terms, abbreviations, citation style, voice, measurement units and recurring product labels. These cannot be solved reliably by isolated segments. The document is the governance unit even when translation is produced sentence by sentence.

Corpus and product ecosystem as unit

Large organisations publish families of documents, interfaces and support pages. Consistency may need to extend beyond one file. A product feature should not have three target names across the website, app and manual. A regulation should use established institutional terminology. A curriculum term should align with official educational usage. At this scale, terminology databases, translation memories, style guides and reference corpora become part of the translation environment.

The communicative event as unit

At the largest scale, a translation can be part of an event: a doctor-patient interaction, a conference, a customer-support exchange, a museum visit, a software task, a court process or a classroom lesson. Text fragments make sense because participants have goals and roles. A one-word utterance such as “Fine” can mean acceptance, quality, penalty or irritation depending on the event.

This does not mean every translation requires ethnographic research. It means the translator should know when the true unit of meaning is larger than the text string.

Translation segmentation in CAT tools

CAT tools segment source text to make translation manageable, align source and target, calculate matches and store reuse. Phrase states that texts are split into smaller segments, usually sentences or titles, and confirmed source-target pairs can be saved into translation memory. memoQ documents join and split operations for cases where sentence segmentation is incorrect or where two source segments should be combined.

These workflows reveal the engineering trade-off. Segments should be small enough for reuse and editing, but large enough to preserve meaningful context. No segmentation rule can solve every language, file format and genre perfectly. Human judgment remains necessary.

Segmentation rules are hypotheses

A segmentation rule may assume that a period followed by a capital letter ends a sentence. Abbreviations, decimals, initials, numbered headings and unusual punctuation can violate that assumption. XML or XLIFF units may already impose boundaries. Tags may interrupt sentences. Lists may contain fragments. Captions may omit punctuation.

Treat automatic segmentation as a first hypothesis about structure. Inspect suspicious breaks before translating.

When to join segments

Join when segmentation divides one grammatical or semantic unit in a way that prevents safe translation. Typical cases include a sentence split at an abbreviation, a term divided by formatting tags, a conditional separated from its result, a quotation broken in a way that obscures syntax, or a source export that placed line breaks inside one sentence.

Before joining, consider downstream effects. Will the merged unit reduce translation-memory reuse? Will tags remain safe? Does the project permit structural changes? Use joining as correction, not as a substitute for reading context.

When to split segments

Split when a source segment contains multiple independent units that need separate handling, reuse or target structure. A long source sentence may contain two complete instructions joined by punctuation. A localisation string may combine a heading and body text because of poor extraction. A bilingual alignment may be wrong.

Again, splitting is not automatically better. Preserve the logical relationship between the resulting pieces and verify the final text in context.

Translation memory and the problem of decontextualised reuse

Translation memory is powerful because repeated source segments can retrieve earlier target segments. The risk is that a stored unit can look self-contained when its original meaning depended on context. A sentence such as “It is now available” may refer to a product, feature, report or service. A target language may require grammatical gender or a more explicit subject. Reusing the old target blindly can create a mismatch.

Context-aware matching, metadata and human review reduce this risk. Phrase documentation describes context-based match behaviour and metadata such as domain, client and filename. The general lesson is broader than any tool: store enough context to make reuse trustworthy.

Termbase versus translation memory

A termbase stores concept-level terminology. Translation memory stores segment-level source-target pairs. Mixing the two can cause confusion. If a term must stay stable across hundreds of sentences, manage it as terminology rather than relying on chance recurrence in translation memory. If a full sentence recurs, memory can reuse the larger pattern.

The correct resource follows the unit of control: concept, segment, style rule or exceptional decision.

AI translation and chunk size

AI systems introduce another version of the unit problem. Users often paste one sentence, one paragraph, a page or an entire document into a model. Larger context can help resolve reference and terminology, but simply increasing input length does not guarantee better translation. Important instructions can become diffuse, local constraints can be overlooked, and output review becomes harder.

The same minimum-sufficient-context rule applies. Give the system enough context to resolve dependencies, plus explicit terminology, audience, register and formatting constraints. Ask it to translate in reviewable units while maintaining a document-level glossary and decision policy.

Context window is not translation unit

An AI model may receive a large context window containing the whole document, but the translation unit can still be one paragraph. This is analogous to a human translator having the full document open while working on one segment. Input context and output unit are separate design choices.

The advantage of separating them is control. Broad context supports interpretation. Smaller output units support verification. A good AI workflow can preserve both.

A robust AI unit workflow

First, provide the brief: source language, target language and locale, audience, purpose, domain, register and risk. Second, provide or derive terminology. Third, give enough surrounding text to resolve local dependencies. Fourth, ask for a bounded target unit. Fifth, compare it to the source. Sixth, carry approved terminology and decisions forward. Seventh, review the assembled target document for cohesion.

This prevents two common AI failures: translating fragments without context and translating huge documents without inspectable decision points.

Machine translation segmentation

Machine translation systems often operate on sentence-like segments or subword sequences internally, but professional users should focus on observable output quality rather than guessing hidden model boundaries. The practical questions remain: did the system preserve relationships across the source, and did segmentation cause omission, duplication, reference drift or unnatural discourse?

Post-editing should therefore include cross-segment review. A sentence can be accurate in isolation and inconsistent with the document.

Quality assurance at unit boundaries

Boundary errors create recognisable quality problems. MQM distinguishes dimensions such as accuracy, terminology, linguistic conventions, style, locale conventions, audience appropriateness and design or markup. Wrong translation units can contribute to several of these at once: an omitted negation becomes an accuracy error; a split technical term becomes a terminology error; a locally grammatical sentence can become unidiomatic in discourse; a broken tag becomes a markup problem.

ISO 5060:2024 provides guidance for evaluation of human, post-edited machine and unedited machine translation output using analytic error approaches. Unit selection matters because evaluators must locate and classify the span in which the problem occurs. Good segmentation makes errors inspectable without pretending that every error is confined to one segment.

Boundary audit 1: omissions at joins

When two source units are merged in translation, small elements can disappear: conjunctions, qualifiers, repeated subjects, exceptions or modal verbs. Review joins explicitly. Compare the target against each source component and verify that no information was absorbed into a smoother but weaker sentence.

Boundary audit 2: additions at splits

When one source unit becomes two target units, translators may add connectors or subjects to make the pieces natural. Some additions are grammatical necessities. Others invent logic. Check whether added “therefore,” “because,” “then,” “we” or “it” is actually supported by the source.

Boundary audit 3: reference drift

Pronouns and noun phrases often shift when sentences are split or reordered. Verify that every “it,” “they,” “this,” “that,” “former” and “latter” still points to the intended referent after translation. If the target language requires repetition of the noun, make the reference explicit without changing emphasis unnecessarily.

Boundary audit 4: terminology drift

Long documents translated in small segments invite synonym variation. In ordinary prose, variation may be natural. In technical terminology, it may imply different concepts. Search the target for central terms after drafting and compare them against the termbase.

Boundary audit 5: tone drift

If each segment is polished independently, the document can alternate between formal and casual, direct and deferential, technical and conversational. Review paragraphs and sections as larger units to stabilise voice.

Boundary audit 6: discourse drift

Check transitions. Does the target preserve contrast, cause, concession, sequence and example structure? Are paragraph openings coherent? Does a conclusion still follow from the evidence? Discourse problems often appear only after segments are assembled.

Boundary audit 7: design and markup

Tags, placeholders, links and formatting can cross segment boundaries. A source may place emphasis tags around part of a phrase. A placeholder may need grammatical agreement. A hyperlink anchor may be split from surrounding text. Review the rendered result, not only the bilingual grid.

Worked example 1: “take off”

Source: “The plane took off at six.” The unit is “took off,” not “took” plus “off.” The phrase means departed into flight. A target language may use one verb. Source: “He took off his coat.” The same visible phrase has a different sense: removed. Source: “The new product took off.” It may mean became rapidly successful. The translation unit is the phrasal verb plus enough clause context to select its sense.

Worked example 2: “current charge”

In an electrical text, “current” and “charge” are technical concepts. In a billing interface, “current charge” may mean the amount presently owed. Translating either word without domain context is unsafe. The visible two-word phrase is not enough unless the section establishes the domain. The understanding unit may therefore be the phrase plus the surrounding screen or document.

Worked example 3: “not uncommon”

“Not uncommon” is a litotic expression. Translating “not” and “uncommon” independently may produce an awkward double negative in the target. The unit is the expression and its rhetorical force: moderately common, or at least more common than one might expect. The target may preserve the understatement or choose a natural equivalent depending on genre.

Worked example 4: “only after”

Source: “Only after the test ended did the error become visible.” The fronted “only after” triggers inversion and highlights timing. The unit includes the temporal phrase and main clause because the emphasis relationship matters. A target language may not use inversion, but it should preserve the restricted timing: the error became visible only once the test had ended.

Worked example 5: conditional warning

Source: “If the seal is damaged, do not use the device.” The condition and prohibition are one functional unit. Translating them separately can weaken the connection. The target must tell the reader exactly when the prohibition applies. In safety text, unit integrity protects action logic.

Worked example 6: two-sentence contrast

Source: “The update installed successfully. However, the service did not restart.” A translator must preserve the contrast between installation success and service failure. “However” cannot be selected sensibly without the first proposition. The understanding unit is the sentence pair even if the production unit remains two sentences.

Worked example 7: pronoun across segments

Source: “The regulator sent the company a notice. It required a response within ten days.” Does “it” refer to the notice or regulator? Most likely the notice, but the target language may force an explicit noun. The unit must include both sentences to resolve reference. If ambiguity remains materially possible, query or preserve it.

Worked example 8: term defined globally

A contract defines “Services” as a specific set of obligations. Later, one segment says “The Services will commence on the Effective Date.” The sentence is locally simple, but the capitalised term has document-wide meaning. The translator must use the established target defined term. The understanding unit includes the definition regime of the contract.

Worked example 9: table header

Source header: “Rate.” In a financial table, it may be interest rate. In a chemistry experiment, reaction rate. In a hotel sheet, nightly rate. In an employee review, rating may even be intended if the source is abbreviated poorly. The cell is one word; the table is the context unit.

Worked example 10: button “Apply”

On a filter panel, “Apply” means activate selected filters. On a job site, it may begin an application. In a graphics editor, it may commit an effect. In a form, it may submit. The target word depends on action. Translate the user interaction, not the isolated label.

Worked example 11: subtitle interruption

Dialogue: “If you think I’m going to—” followed by another speaker: “I know.” The first subtitle is incomplete by design. Completing it in translation would invent speech. The unit includes turn-taking and interruption. Subtitle segmentation must respect dialogue function as well as reading speed.

Worked example 12: joke setup and payoff

A joke may place its setup in one sentence and payoff in the next. Translating each independently can erase wordplay or reveal the punchline too early. The joke is the understanding unit. Target wording may require coordinated choices across both sentences, even if subtitle or page layout keeps them separate.

Worked example 13: academic hedge

Source: “The pattern may indicate a modest association. Further work is needed.” The two sentences jointly communicate caution. Translating the first strongly and the second literally can create inconsistency. The paragraph-level stance is the unit for review.

Worked example 14: legal exception

Source: “All records must be retained for seven years, except records listed in Schedule B.” The exception is part of the obligation. If segmentation separates the exception, a reviewer may miss it or a memory match may reuse only the rule. Treat rule plus exception as one legal reasoning unit.

Worked example 15: medical dosage

Source: “Take one tablet twice daily with food.” The dosage unit combines quantity per dose, frequency and condition. “Two tablets daily” preserves total tablets but changes schedule. Do not split numerical pieces from their relation.

Worked example 16: technical sequence

Source: “After saving the configuration, restart the controller and wait until the status light turns green.” The action sequence is one procedural unit: save, restart, wait for condition. A target that reorders for style can create operational error. Translate the sequence as a process graph before wording it.

Worked example 17: literary motif

A novel repeats the word “door” across several chapters as a motif of exclusion and possibility. Locally, synonyms might sound elegant. Globally, variation can erase the motif. The document or thematic arc becomes the unit for lexical consistency. Literary translation often has long-range dependencies invisible to CAT segmentation.

Worked example 18: character voice

A child narrator uses short clauses, repetition and basic vocabulary across a scene. Polishing each sentence independently into sophisticated target prose changes character. The scene or character voice profile is the unit for stylistic control.

Worked example 19: SEO heading

Source: “How to fix a slow laptop.” A literal target may be understandable but fail to match target-language search phrasing. SEO translation needs a unit larger than the heading: page intent, target queries and actual content. The target heading should express the same informational job using natural search language.

Worked example 20: form label plus help text

Label: “ID.” Help text: “Enter the number shown on your membership card.” The label alone is ambiguous. The help text resolves it. The understanding unit includes both. The production unit may remain a short label and a separate sentence.

Worked example 21: “free”

“Free” can mean no cost, available, unoccupied, released, without restriction or politically free. A pricing page, hotel room status, memory allocation screen and human-rights text activate different senses. The word is not a safe unit until the surrounding task identifies the concept.

Worked example 22: “right”

“Right” can indicate correctness, direction, entitlement, immediately or moral justification. “Turn right” and “human rights” have no shared target equivalent in many languages. The phrase or domain defines the unit.

Worked example 23: headline compression

Headline: “Minister backs school reform critics oppose.” English headline grammar omits articles and can create attachment ambiguity. The translator should expand the likely proposition from the article before translating. The headline is not an autonomous unit if its parse depends on body text.

Worked example 24: quotation attribution

News text: “The measure is unnecessary,” the minister said, adding that implementation would be delayed. The quote, attribution and added clause distinguish speaker claim from reporter narration. Splitting them carelessly can blur ownership of the statement. The reporting unit includes attribution.

Worked example 25: “since”

“Since Monday” is temporal. “Since the door is locked, we will wait” is causal. The connective cannot be translated before its clause relation is identified. The translation unit is the connective plus proposition structure.

Worked example 26: “while”

“While she cooked, he read” is temporal. “While the method is fast, it is not reliable” is concessive. Same word, different discourse relation. Clause-level analysis selects the target.

Worked example 27: “as”

“As the temperature rose, pressure increased” may encode temporal correlation, causal interpretation or both. “As a teacher” marks role. “As requested” marks conformity. Translate the construction, not the token.

Worked example 28: “would”

“Would” can mark conditionality, habitual past, future-in-the-past, politeness or reported intention. “She said she would come” differs from “Would you help?” and “Every summer we would visit.” The clause and discourse function define the unit.

Worked example 29: “still”

“Still working” implies continuation. “Still, we proceeded” is a discourse marker meaning nevertheless. “Stand still” means not moving. Translate the grammatical construction and local relation, not the spelling.

Worked example 30: cross-sentence ellipsis

Question: “Who approved the change?” Answer: “Maria.” The one-word answer contains an implicit proposition: Maria approved the change. A target language may express the answer differently depending on grammar or politeness. The question-answer pair is the understanding unit.

Worked example 31: repeated terminology with changing sense

A document uses “model” first for a statistical model and later for a product model number. Blind consistency produces error. Terminology consistency applies to concepts, not strings. The translator must classify each occurrence by sense before enforcing sameness.

Worked example 32: repeated ordinary word with one technical occurrence

A biology text uses “cell” mainly as a biological term, but one table mentions a spreadsheet cell. The string is identical; the concepts differ. A termbase can separate domains or definitions. The local unit selects which concept applies.

Worked example 33: cross-reference

“See Section 4 for exceptions.” The phrase is simple, but the target must preserve reference numbering, document structure and possibly translated heading names. The unit includes the navigation function. If sections are renumbered in localisation, automated cross-reference handling may be safer than manual text changes.

Worked example 34: placeholder grammar

String: “{count} files deleted.” The placeholder may represent one or many. English plural handling differs from other languages. The unit is not the static string alone; it includes variable behaviour and plural rules. Localisation often needs plural-aware message structures rather than one translated sentence.

Worked example 35: gendered placeholder

“{name} updated their profile.” English can use singular they. A target language may require grammatical gender or a neutral restructuring. The translator needs product knowledge about available user data and inclusivity policy. The unit includes the dynamic system.

Worked example 36: date string

“03/04/2026” cannot be safely translated without locale. The visible characters are one string, but the meaning unit includes source date convention. A target may need “3 April 2026” or “4 March 2026.” Locale metadata is part of the evidence.

Worked example 37: decimal formatting

“1,500” can mean one thousand five hundred in one convention and one point five in another context using comma decimals. Numbers require locale and domain context. The unit includes numeric convention, not only digits.

Worked example 38: measurement unit

“5 ft” may need preservation, conversion or dual display depending on the brief. Converting to metres changes the text but can improve usability. Conversion is not a language-only decision. The unit includes measurement policy and target audience.

Worked example 39: product name plus common noun

A brand name may look like an ordinary word. Translating it can destroy identity. Verify whether the expression is a registered product, a feature name or a descriptive phrase. Proper-name policy can make the difference between translation and accidental renaming.

Worked example 40: institutional title

An organisation may publish an official target-language name. Use it even if a literal translation seems possible. The unit extends beyond the source string to institutional naming conventions.

Translation units in literary work

Literary translation often requires larger units because effects accumulate. Rhythm can span a sentence. Imagery can extend across a paragraph. Repetition can become a motif across a chapter. Character voice can persist throughout a book. A pun can rely on an earlier phrase. The translator may draft line by line, but must keep these long-range patterns active.

The danger of small units in literature is local elegance at the expense of global design. The danger of very large units is free rewriting. The discipline is to identify which formal features are invariant for the passage and coordinate local choices around them.

Translation units in technical work

Technical translation often benefits from smaller, stable production units because procedures, component names and warnings must remain traceable. Yet terminology definitions, diagrams and step sequences create larger understanding units. A component label may be one word; its meaning may come from an engineering schematic. A warning may govern several numbered steps.

Technical translators should combine segment-level precision with document-level terminology governance.

Translation units in legal work

Legal translation makes unit boundaries consequential. Defined terms have document-wide reach. Conditions and exceptions must remain attached to the rule they qualify. Cross-references link clauses. A sentence can be long because legal drafting packs relationships into one provision. Splitting for readability may alter scope if done carelessly.

The translator should diagram legal relations before restructuring. Rights, duties, permissions, prohibitions, conditions, parties and temporal limits are semantic units even when grammar changes.

Translation units in medical work

Medical translation combines terminology, dosage, risk, sequence and audience. A medicine instruction may be one short sentence but contain several inseparable variables. A patient information leaflet may use a specialist term followed by a plain-language explanation; translating them as unrelated segments can break the teaching function.

Review medical units by action and consequence, not only sentence boundary.

Translation units in academic work

Academic argument often spans sentences. Claims, evidence, limitations and hedges work together. A paragraph can move from observation to cautious interpretation to limitation. Translating each sentence at a different level of certainty distorts the research stance. Citations also define claim ownership and should remain attached to the propositions they support.

Translation units in business communication

Business messages often combine relationship management with action. An email may open warmly, explain a constraint, make a request and propose a deadline. Translating only the request can make it sound harsher than the whole message. Translating only the politeness formulas can make the action unclear. The email is a functional unit even if production occurs paragraph by paragraph.

Translation units in marketing and transcreation

Marketing copy can distribute one promise across headline, subheading, visual and call to action. A headline may be deliberately incomplete until the image supplies context. The unit is the campaign element, not necessarily the sentence. Transcreation can therefore require larger multimodal units while remaining accountable to the product claim.

Translation units in websites

Web pages combine headings, navigation, metadata, body copy, buttons, alt text and structured data. Each has different constraints, yet users experience them together. Translators need page-level intent and component-level production. SEO keywords may apply to title and heading, while terminology must remain consistent in body and interface.

Translation units in apps

Apps are especially hostile to context because strings can be extracted from code and sorted alphabetically. The translator sees “Next,” “Done,” “Open” and “Remove” without state information. Add string keys, screenshots, developer notes and user-flow context. A one-word production unit may need a whole interaction as its understanding unit.

Translation units in subtitles

Subtitle segmentation must coordinate spoken meaning, reading speed, timing, shot changes and turn-taking. A clause may be split across subtitle cues for timing, but the semantic unit remains intact. Conversely, two short utterances may share one cue. Translators should avoid breaking tight syntactic groups in places that slow comprehension.

Translation units in interpreting

Interpreting makes units temporal. Consecutive interpreters may work with idea units that fit memory and notes. Simultaneous interpreters work with short meaning groups while predicting sentence structure. The written translation unit model still helps: do not commit before enough evidence is available to identify the proposition, but do not wait so long that working memory collapses.

Translation units in education

Students often translate one word at a time because vocabulary exercises train isolated recall. Teaching the unit ladder changes the habit. Ask learners first whether the unknown item is a word, phrase, clause or sentence relation. Let them mark collocations, idioms, phrasal verbs, pronouns and connectives. Then translate.

The objective is not to ban dictionaries. It is to use dictionaries after deciding what unit needs a meaning.

A Primary-level unit exercise

Give learners short pairs such as “take a photo,” “take a bus,” “take a break” and “take medicine.” Ask whether one target word for “take” works in every case. The lesson is that meaning often belongs to the phrase. Then use a picture or short scene to supply context. Children can discover translation units without technical terminology.

A Secondary-level unit exercise

Give students a paragraph containing one idiom, one phrasal verb, one pronoun whose referent is in the previous sentence, one connective and one ambiguous word. Ask them to draw brackets around the smallest unit needed to solve each item. They will see that different problems require different bracket sizes.

An advanced unit exercise

Use a CAT-style table with sentence segments. Hide the paragraph at first and ask students to translate. Then reveal the preceding and following sentences and let them revise. Discuss which changes occurred because the initial unit was too small. This teaches context sensitivity and translation-memory caution.

The bracket method

A simple manual technique is to place mental brackets around dependencies. In “The results were not statistically significant,” the phrase “not statistically significant” is a technical evaluation unit. In “not all results were significant,” the bracket must show that negation scopes over “all.” In “the system that failed yesterday,” the relative clause belongs with the noun it modifies.

Bracketing makes invisible structure visible before translation.

The paraphrase test

If you are unsure whether the unit is large enough, paraphrase the source in the source language. If the paraphrase needs words from outside the current segment to be accurate, the understanding unit is too small. If the paraphrase can state the full relation cleanly, you probably have enough context.

The substitution test

Replace one suspected word with a synonym. If the surrounding phrase becomes impossible or changes meaning sharply, the expression may be collocational or fixed. “Heavy rain” accepts some variation, but “strong rain” sounds wrong in ordinary English. The natural target should be chosen at the collocation level.

The pronoun test

Circle every pronoun and demonstrative. Can you name the referent from inside the current unit? If not, expand. This catches one of the most common cross-segment failures in translation and machine output.

The connector test

For every however, therefore, because, although, instead, meanwhile, still and similarly, state the relation in plain language. If you need another sentence to describe the relation, include that sentence in the understanding unit.

The term test

Ask whether the candidate expression is an ordinary word or a domain term. Search definitions, glossary resources, standards and authoritative target-domain usage. If the term is defined elsewhere in the document, the definition becomes part of the evidence even if it is many pages away.

The layout test

Ask whether visual position changes meaning. A word inside a button, warning, chart, label or menu is not fully specified by text alone. Preview the page or screen. If layout determines function, include it in the understanding unit.

The risk test

Higher consequence justifies larger context and stricter review. A casual message can tolerate some uncertainty. A dosage, safety warning, legal deadline or financial condition should be checked against the widest relevant context. Unit size is partly a risk-management decision.

The reusability test

Before storing a translation-memory unit, ask whether it can safely recur elsewhere. If the translation depends on a specific gendered referent, product, legal definition or screen state, preserve context metadata or avoid treating the target as universally reusable. Reusability is a property of a source-context pair, not merely a source string.

The naturalness test

Read the target without seeing segment boundaries. Does it flow as target-language discourse? If the sentence transitions, pronouns or repeated terms sound mechanical, the production units may have been handled too independently. Larger review units repair this.

The fidelity test

Then restore the source and compare. Larger target units can tempt elegant paraphrase that omits qualifiers. Verify every actor, action, condition, exception, quantity, modal and relation. Naturalness and fidelity must survive together.

Diagnostic: fluent sentence, wrong unit

If a translation is fluent but wrong, ask whether the translator solved the wrong-sized problem. A word may have been interpreted without its phrase, a sentence without its paragraph, a button without its screen, a claim without its citation, a term without its definition. Unit diagnosis often explains why skilled target-language writing can still be inaccurate.

Diagnostic: awkward sentence, right meaning

If meaning is correct but the target sounds translated, the production unit may be too small. Source word order or phrase boundaries may have been preserved when the target language prefers a different clause or sentence structure. Rebuild a larger target unit while protecting the semantic invariants.

Diagnostic: inconsistent terms

If the same concept receives several translations, the document governance unit was ignored. Move from segment-level decision making to termbase control. If the same string legitimately represents different concepts, do the reverse: stop enforcing blind string consistency and classify senses.

Diagnostic: repeated pronoun errors

Repeated pronoun mistakes indicate that the context window is too narrow. Expand by at least one or two preceding sentences and build a reference chain. For AI translation, include previous translated context or entity notes.

Diagnostic: sentence-by-sentence tone wobble

If one paragraph alternates between formal and casual wording, review at paragraph and document level. Establish address form, terminology and brand voice as global constraints.

Diagnostic: translation memory keeps suggesting the wrong answer

Investigate whether old units lack context metadata, whether the domain has changed, whether the source string is polysemous, or whether different target versions were overwritten. Translation memory quality depends on the segmentation and governance decisions that created it.

Diagnostic: machine translation repeats or omits material

Check segmentation and input packaging. Broken tags, duplicated segments, long blocks with repeated structures or truncated context can contribute. Compare the source and target at multiple scales: segment, paragraph and document. Do not assume the error is purely lexical.

Diagnostic: one-word interface translations keep changing

Add string keys, screenshots and action descriptions. A one-word string can legitimately require different translations across functions. Consistency should be based on function, not spelling alone.

Diagnostic: subtitles read badly even when accurate

Check whether syntactic groups are split across cues, whether pronouns are separated from antecedents, whether reading time forces compression and whether speaker turns are clear. Subtitle units are audiovisual, not merely textual.

Diagnostic: legal clause becomes clearer but less precise

Restructuring may have moved an exception, weakened a modal or changed which phrase a condition modifies. Reconstruct the legal relationship before simplifying target syntax. Clarity is desirable only when legal scope remains stable.

Diagnostic: literary translation loses rhythm

The line or sentence may be too small. Read the passage aloud. Track repetition, sentence length, pauses, parallelism and image patterns across the paragraph or stanza. Recreate the larger formal unit.

Vocabulary and the unit problem

Vocabulary learning often begins with word-definition pairs. Translation reveals the limits of that model. A word’s meaning can depend on collocation, grammatical construction, domain and discourse. Good translators need vocabulary depth: senses, connotations, register, word families, collocations and constraints.

The protected Vocabulary Learning Hub remains the owner of systematic vocabulary learning. This translation node uses that knowledge to answer a different question: at what span should lexical knowledge be applied? A vocabulary weakness may make a unit hard to interpret, but enlarging the unit can also make the vocabulary easier by adding context.

English and the unit problem

English grammar creates many dependencies that are larger than individual words: phrasal verbs, noun compounds, relative clauses, tense-aspect sequences, modal scope, inversion, articles, ellipsis and information structure. A translator who knows the words but cannot parse these structures may choose the wrong unit.

How English Works remains the owner of those mechanisms. The translation architecture uses them to decide how much English must be understood together before movement into another language.

Language pairs and unit asymmetry

A translation unit need not have the same surface size in both languages. One source word can become a target phrase. A source phrase can become a target word. A source sentence can become two target sentences. Two short source sentences can become one natural target sentence. Unit asymmetry is normal because languages package information differently.

The goal is not structural mirroring. The goal is preservation of the required semantic and communicative relationships.

English and Chinese unit shifts

English may require explicit articles, tense choices and subjects that Chinese handles differently through context, aspect markers, classifiers or topic structure. A Chinese source can omit a subject across several clauses, forcing an English translator to infer who acts. An English noun phrase may require a different structural packaging in Chinese. The translator often needs a paragraph-level understanding unit even when producing shorter target clauses.

English and Japanese unit shifts

Japanese frequently omits recoverable subjects and encodes social relationships through politeness and honorific choices. English can leave some of those relationships less explicit. Translation may therefore require discourse context larger than the sentence. A target English pronoun can be impossible to choose safely without identifying the speaker, addressee and topic chain.

English and Romance-language unit shifts

English phrasal verbs, noun compounds and flexible conversion between noun and verb may become more explicit phrases in French, Spanish, Italian or Portuguese. Conversely, target morphology may compact information. Cognates can tempt word-level translation even when collocation or register differs. Treat resemblance as a candidate, then verify the phrase-level unit.

English and Malay unit shifts

English tense marking, articles and compact noun phrases do not map mechanically to Malay. Context may carry temporal information; affixation and reduplication can package meaning differently. A translator should identify proposition and discourse function before deciding how many target words are needed.

English and Arabic unit shifts

Arabic morphology can package information that English spreads across words, while register and regional variation affect target choices. Right-to-left layout creates additional interface considerations. Names and transliteration policies may operate at document level. The unit problem therefore combines language, product context and visual design.

Unit size and translation speed

Beginners often think smaller units are faster. They can be slower because every word becomes a separate decision. Skilled translators recognise familiar multiword patterns and translate them as units. At the same time, working with excessively large units can slow review and reduce reuse. Speed comes from selecting the right unit quickly.

Pattern recognition is therefore a speed skill. It is also an accuracy skill.

Unit size and cognitive load

Working memory is limited. Translators must hold source meaning, target alternatives, terminology, grammar and context simultaneously. Breaking work into units reduces cognitive load. The art is to chunk without severing dependencies. This is why experienced translators alternate between local focus and global scanning.

Unit size and revision

Different revision passes need different units. Meaning review can work clause by clause. Terminology review may be document-wide. Tone review may work paragraph by paragraph. Layout review must happen in the final interface. Numbers can be checked with automated comparison. No single unit is optimal for every quality task.

The layered review model

Pass 1: clause and sentence accuracy. Verify propositions, actors, conditions, negation, numbers and modal force.

Pass 2: paragraph cohesion. Verify reference, transitions, information flow and tone.

Pass 3: document consistency. Verify terminology, names, style rules, locale conventions and recurring phrases.

Pass 4: in-context function. Verify interface behaviour, layout, subtitle timing, links, tables and reader tasks.

The segmentation decision log

For difficult projects, record unusual unit decisions. Example: joined segments 54 and 55 because the source export split a conditional clause; kept product slogan as one unit to preserve wordplay; treated UI string key checkout.submit as a distinct sense of “Submit”; used paragraph context to resolve pronoun in segment 118.

This record helps reviewers understand why the translation departs from default segmentation and helps teams repeat good decisions.

Team translation and shared units

When several translators work on one document, segment assignment can hide global dependencies. Give teams the full brief, termbase, style guide, entity list and access to surrounding context. Avoid assigning interdependent paragraphs to different translators without coordination. Review seams between translators as carefully as seams between segments.

Parallel texts as larger units of evidence

Sometimes the correct target expression cannot be chosen from the source alone. Comparable target-language documents reveal genre conventions, terminology and phraseology. A technical manual, research article, government form or museum label has a conventional target style. The evidence unit becomes a small corpus of authentic target texts.

Corpora and phrase-level evidence

Corpora are particularly useful for collocation and register. Search candidate phrases rather than isolated words. Ask which nouns occur with the verb, which adjectives modify the term, which preposition follows the expression and which genre uses it. Translation research improves when queries match the suspected unit.

Dictionaries and unit awareness

Dictionaries remain essential, but use them at the right scale. Monolingual dictionaries help identify source sense. Bilingual dictionaries generate target candidates. Phraseological dictionaries and corpora reveal multiword patterns. Specialist glossaries stabilise terms. A dictionary entry is evidence, not an instruction to replace one token with another.

Terminology databases and concept units

A termbase is concept-oriented. The same concept can have multiple surface forms, abbreviations and grammatical variants. The same string can represent different concepts. Concept management therefore sits above simple segment matching. It is especially important in technical, legal, medical and institutional translation.

Unit choice in back-translation

Back-translation can reveal drift, but unit size matters. If the target naturally restructures a sentence, back-translating tiny phrases may make a good translation appear different. Compare propositions and functional units, not only word alignment. The question is whether required meaning survived, not whether the surface can be reversed token by token.

Unit choice in bilingual review

A bilingual reviewer should move between scales. Align source and target closely enough to catch omissions and additions, then zoom out to paragraph and document levels for coherence. Reviewers who stay only at sentence level can miss tone and terminology systems. Reviewers who stay only at document level can miss a missing negation.

Unit choice in automated QA

Automated QA excels at repeatable features: numbers, terminology, tags, punctuation, missing target segments, repeated words and formatting. These checks often operate at segment or document scale. Human review is still needed for dependencies such as reference, pragmatic force and context-sensitive sense.

MQM and unit-aware quality diagnosis

MQM provides a structured error typology including accuracy, terminology, linguistic conventions, style, locale conventions, audience appropriateness and design or markup. A unit-aware reviewer can use these categories to identify not just what is wrong but why. A false friend may reflect word-level overconfidence. An omission may occur at a join. Inconsistent terminology may reflect overly local decision making. Unidiomatic style may reflect source-segment syntax carried into the target.

ISO 5060 and evaluation scope

ISO 5060:2024 addresses evaluation of human translation, post-edited machine translation and unedited machine translation output, with an analytic approach based on error types and penalty points. It also notes that its guidance can support evaluation of source texts intended for translation. Unit selection fits naturally into this framework because evaluators need to decide what span contains an error and what surrounding evidence determines severity.

ISO 17100 and process discipline

ISO 17100:2015 remains the published international standard for translation-service requirements while a new edition is under development as of 2026. Its emphasis on core processes, resources and applicable specifications reinforces a broader lesson: translation quality is produced by a controlled process, not only by individual sentence skill. Unit governance belongs inside that process through briefs, terminology resources, revision and review.

How to choose unit size under time pressure

Use a two-speed rule. For green segments with stable sense and low risk, translate at the default sentence or phrase level. For amber segments containing new terminology, unusual syntax, pronouns, idioms or cross-sentence links, expand context. For red segments involving safety, law, medicine, financial values, unresolved ambiguity or critical user actions, use the widest relevant context and require deliberate verification.

The 30-second Unit Gate

Ask four questions: What is the smallest span that expresses the proposition? What outside information could change the translation? Is any term or reference governed elsewhere? What will the reader do with this text? If the answers fit inside the current segment, proceed. If not, expand.

The five-minute unit map

For a longer document, scan headings, definitions, repeated terms, entities, warnings, tables and interface notes. Mark places where segment boundaries are likely to be misleading. Record global style and terminology decisions. This map reduces the number of times you discover context too late.

The project-level unit architecture

Define four layers before large-scale translation. Layer 1: terminology units managed in the termbase. Layer 2: segment units managed in CAT or translation memory. Layer 3: discourse units reviewed by paragraph, section and document. Layer 4: product-context units reviewed in interface, layout or real-world use. Quality improves when each layer has an owner and review method.

Source preparation can improve units

Writers who know content will be translated can reduce avoidable unit problems. Use consistent terminology, clear pronoun reference, stable headings, complete sentences where appropriate, meaningful string keys, translator notes, well-formed markup and explicit variables. Avoid splitting one sentence across unrelated software strings. Avoid placing multiple unrelated meanings in one translatable string.

Better source architecture produces better translation architecture.

Do not over-control creative source texts

Controlled-language principles help operational content but can damage literature, dialogue and branding if applied mechanically. Deliberate ambiguity, fragmentation, repetition and unusual syntax may be features rather than defects. The unit decision should preserve purposeful form, not normalise it away.

Unit choice and ambiguity

Expanding the unit often resolves ambiguity, but not always. Some sources remain genuinely ambiguous at document level. Do not keep expanding forever. Once all relevant evidence is inside the envelope, classify the ambiguity as unresolved. Preserve it when possible, query when consequential, or record the interpretive choice when the target forces resolution.

Unit choice and uncertainty

Mark uncertainty before drafting. A translator can tag TERM?, REFERENT?, SCOPE?, IDIOM?, DATE?, UNIT?, CULTURE? or UI? and then resolve the issue at the appropriate scale. Fluent target language can hide uncertainty, so the unit-analysis stage is the right place to make it visible.

Unit choice and risk

The cost of a wrong boundary varies. A literal idiom in a casual travel blog may be embarrassing. A split negation in a medicine instruction can be harmful. A misplaced exception in a contract can change obligations. Risk should therefore influence how aggressively translators expand context, research terminology and request review.

Unit choice and reader effort

A target text should distribute information in units that target readers can process naturally. Source sentences can sometimes be too long for the target genre; short source fragments can sometimes need merging. The translator must preserve meaning while designing readable target units.

Unit choice and accessibility

Accessible translation may require shorter sentences, explicit headings or clearer link text, but restructuring must not disconnect conditions, labels and explanations. Screen readers and captions create additional unit constraints. A link labelled “click here” may be less useful than a descriptive target. Accessibility is a functional translation condition.

Unit choice and multilingual design

Target languages vary in text expansion, word length, line breaking and grammatical needs. A button width designed around an English one-word label may fail in another language. The UI component and its spatial constraints become part of the unit. Internationalisation should allow target-language units to breathe.

Unit choice and localisation

Localisation extends beyond text. Dates, numbers, currency, measurement, address fields, images, legal notices and cultural expectations can be unit-level dependencies. A string may be linguistically correct but functionally wrong because the surrounding locale logic remains unchanged.

Unit choice and transcreation

Transcreation often enlarges the unit to a campaign concept, slogan-plus-visual or audience response. This does not remove accountability. The translator still identifies the invariant: product promise, emotional effect, call to action, brand personality or legal constraint. The target may diverge lexically while preserving the higher-level unit.

Unit choice and names

Names can have established target forms, transliterations, translations or protected spellings. A person’s name may be one token but require document-level consistency. An organisation’s legal name may include words that should not be translated freely. Maintain an entity list.

Unit choice and citations

Academic and journalistic texts attach claims to sources. A citation placed at the end of a sentence may support only one clause or a whole paragraph. Translation should preserve claim-source boundaries. Moving citations for target style can accidentally change attribution.

Unit choice and quotations

A quotation can contain its own voice, historical register and legal status. Decide whether an official existing translation should be used. If the quote is translated anew, keep quotation boundaries and attribution clear. The surrounding paragraph may explain how the quotation functions.

Unit choice and headings

Headings often omit articles and verbs. Their meaning may depend on the section below. Translate headings after understanding section purpose, not before. A heading can also establish terminology reused in navigation and metadata, so the decision may need product-wide consistency.

Unit choice and captions

Captions are multimodal. “Left: control specimen” makes no sense without the image. The unit includes visual layout and labels. In scientific and museum contexts, caption accuracy can depend on figure numbering, orientation and domain terminology.

Unit choice and footnotes

Footnotes may clarify terminology, cite evidence or add optional commentary. The main text and note can form one interpretive unit. Translating the note independently can produce repeated explanations or inconsistent terms. Review note and anchor together.

Unit choice and lists

List items may be fragments governed by a shared lead-in. If the lead-in says “The system can:” and items begin with verbs, each item inherits grammar from the lead-in. Translating items in isolation can break parallelism. The list plus lead-in is the grammatical unit.

Unit choice and tables

Table cells inherit meaning from row and column headers. “High” can refer to risk, temperature, priority or score. The cell is not self-contained. Translate tables by reading axes, units, notes and surrounding explanation.

Unit choice and diagrams

Diagram labels can be extremely short, but arrows and spatial relationships supply syntax. A translator should inspect the diagram before choosing target labels. Text expansion may also require layout changes.

Unit choice and code

Code, variables and tags create hard boundaries. The translator must know which tokens are translatable and which are protected. A placeholder may sit inside a grammatically larger unit. Never translate a key simply because it resembles ordinary language unless the localisation framework marks it as user-facing.

Unit choice and structured content

CMS content can separate headline, excerpt, body, metadata and call-to-action fields. They may be stored independently but appear together. The translator needs page-level context while producing field-level target text. Structured content makes the understanding-production distinction especially important.

Unit choice and email threads

An email reply such as “That works for us” depends on the proposal in the previous message. Translating only the reply can leave “that” ambiguous. In multilingual support or legal discovery, thread context may be necessary to preserve reference and intent.

Unit choice and chat

Chat messages are short but context-dependent. Emoji, reply threading, quoted messages and timing affect meaning. “Sure.” can be enthusiastic, neutral or reluctant. The conversational turn sequence is often the understanding unit.

Unit choice and social media

Posts can depend on images, hashtags, quoted posts and platform conventions. Hashtags may be searchable labels rather than ordinary phrases. Character limits may force compression. Translate the communicative package, not only caption text.

Unit choice and search snippets

SEO titles and descriptions need page intent, target query language and length constraints. A literal sentence translation may be poor search copy. The page is the semantic unit; the snippet is a constrained production unit.

Unit choice and voice interfaces

Voice commands depend on intent classification, wake words, device state and expected response. A phrase such as “turn it down” requires reference to a device or volume. Translation should align with the command grammar the product recognises. The unit includes interaction logic.

Unit choice and generative AI interfaces

AI assistants may generate dynamic responses rather than fixed strings. Translation can occur through multilingual generation, post-processing or prompt localisation. Here the unit can be a prompt template plus variables, safety instructions, system behaviour and conversation history. Testing must cover generated examples, not only static text.

Unit choice and evaluation datasets

Benchmarking translation on isolated sentences can miss document context. Conversely, document-level evaluation can make error localisation difficult. Good evaluation designs include enough context for fair interpretation and enough segmentation for reproducible annotation. This mirrors the three-unit model.

Unit choice and human agency

Tools suggest boundaries, matches and candidate translations. Humans decide when those suggestions are appropriate. Human agency in translation is not the refusal of automation. It is the ability to recognise when the problem has been framed at the wrong scale and to change the frame.

A 12-step unit workflow for professional translation

1. Read the brief and identify audience, purpose, locale and risk.

2. Scan the document structure and final presentation.

3. Build terminology and entity resources.

4. Inspect automatic segmentation for obvious errors.

5. For each difficult segment, identify the translation decision.

6. Expand the dependency envelope until the decision is evidence-complete.

7. Draft in a manageable production unit.

8. Confirm terminology and reference before storing reusable units.

9. Review paragraph cohesion and section purpose.

10. Run document-level consistency and automated QA.

11. Review in final context.

12. Feed approved decisions back into termbases, memories, style guides and decision logs.

Fifty quick unit diagnostics

1. Does the source expression have a meaning not predictable from its separate words? Expand to phrase.

2. Does a particle change the verb meaning? Treat the phrasal verb as one unit.

3. Does a noun compound hide relationships? Expand the compound before translating.

4. Does negation scope over a quantifier? Use the full clause.

5. Does “only” modify one element? Keep the focus relation inside the unit.

6. Does the sentence contain an if-clause? Translate condition and result together.

7. Does a pronoun point backwards? Include its antecedent.

8. Does “this” refer to a previous proposition? Include the previous sentence.

9. Does a connector express contrast or cause? Include both propositions.

10. Does an abbreviation contain punctuation? Check segmentation before trusting the split.

11. Does a term have a definition? Include the definition as governing context.

12. Does a capitalised expression function as a defined term? Enforce document-wide consistency.

13. Does a repeated word change sense? Do not enforce blind consistency.

14. Does a joke depend on setup? Translate setup and payoff together.

15. Does a metaphor continue across sentences? Treat the metaphor chain as one system.

16. Does a literary motif recur? Track it across the work.

17. Does a character have a distinctive voice? Review at scene or character level.

18. Does a table cell depend on a header? Translate the table context.

19. Does a label depend on an icon? Include visual evidence.

20. Does a button cause an action? Translate the action, not the label alone.

21. Does a placeholder vary by number? Include plural behaviour.

22. Does a placeholder vary by gender or role? Include user-model constraints.

23. Does a subtitle cue cut a clause? Preserve the semantic unit across timing boundaries.

24. Does an interrupted line end unfinished? Preserve interruption.

25. Does an email reply contain “that,” “it” or “this”? Include the earlier message.

26. Does a legal exception follow the rule? Keep them in one reasoning unit.

27. Does a dosage include quantity and frequency? Keep them together.

28. Does a procedure require sequence? Translate the action chain.

29. Does a research claim include hedges? Review claim and limitation together.

30. Does a citation define who owns the claim? Keep attribution clear.

31. Does a marketing slogan depend on an image? Treat the creative as multimodal.

32. Does an SEO heading depend on page intent? Research target query language at page level.

33. Does a product name look translatable? Verify official naming before changing it.

34. Does an institution have an official target name? Use the authoritative form.

35. Does a date use ambiguous numerals? Include source locale.

36. Does a number use locale-specific separators? Verify value before formatting.

37. Does a measurement need conversion? Follow project policy, not habit.

38. Does a translation-memory match come from another domain? Treat it as a candidate, not authority.

39. Does the same source have multiple target possibilities in memory? Compare context metadata.

40. Does a segment feel grammatically incomplete? Inspect neighbouring segments.

41. Does a paragraph feel choppy after accurate sentence translation? Review discourse as a larger unit.

42. Does a document wobble in formality? Apply document-level style control.

43. Does target terminology vary unintentionally? Search the whole document.

44. Does a source segment contain unrelated sentences? Consider splitting for control.

45. Did formatting divide one sentence? Consider joining or using broader context.

46. Is the source intentionally ambiguous? Do not enlarge the unit into invented certainty.

47. Is the source itself defective? Mark the defect rather than repairing silently.

48. Does a term choice require specialist knowledge? Expand to domain evidence.

49. Does the target require a distinction absent in the source? Seek context before guessing.

50. Can you explain why this boundary is safe? If not, run the Unit Gate again.

Frequently asked question: should I translate sentence by sentence?

Sentence by sentence is a strong default production method, especially in CAT tools, but not a complete understanding method. Read surrounding context and expand whenever reference, terminology, tone or logic crosses the boundary. The sentence is a workstation, not an island.

Frequently asked question: is the translation unit always a phrase?

No. It can be a word, phrase, clause, sentence, paragraph, document feature or communicative event. The right unit is whichever span contains the dependencies needed for the current translation decision.

Frequently asked question: what is the difference between a segment and a translation unit?

In CAT-tool practice, the terms can overlap because a confirmed source-target segment is often stored as a translation unit in memory. In reasoning, it is useful to distinguish them: a segment is the chunk presented or stored by the system; the translation unit is the span that must be interpreted together. Sometimes they are identical. Sometimes they are not.

Frequently asked question: should I join CAT-tool segments?

Join when the segmentation is genuinely wrong or when a meaning-bearing structure has been split in a way that blocks accurate translation, and when project rules permit it. Do not join simply to make the screen look comfortable. Consider translation-memory reuse, tags and downstream workflow.

Frequently asked question: should I split long sentences?

You can split them cognitively for analysis and sometimes structurally in the target if target-language style benefits, but preserve logical relations. If the source sentence contains one legal, technical or rhetorical unit, careless splitting can alter scope or emphasis.

Frequently asked question: does more context always improve AI translation?

Not automatically. Relevant context helps. Unstructured volume can make review harder and dilute instructions. Give enough context to resolve dependencies, make terminology explicit and request output in reviewable units.

Frequently asked question: what should go into translation memory?

Store approved source-target segments that are useful for future reuse, with appropriate context and metadata. Manage concepts in a termbase, global writing policy in a style guide and exceptional decisions in a log. Do not force every kind of knowledge into segment memory.

Frequently asked question: why do exact matches sometimes fail?

The same source string can mean different things in different contexts. Domain, surrounding segments, product state, speaker, referent and locale can change the correct target. Exact character identity is not exact meaning identity.

Frequently asked question: how does this connect to vocabulary?

Vocabulary knowledge helps identify senses, collocations, multiword terms, register and semantic distinctions. The translation-unit method tells you where that vocabulary knowledge must be applied. If the problem is learning the word itself, use the Vocabulary Learning Hub. If the problem is deciding what must be translated together, stay in this translation architecture.

Frequently asked question: how does this connect to English?

English syntax and discourse determine many unit boundaries: phrasal verbs, noun compounds, clauses, relative structures, inversion, ellipsis, tense, aspect and information structure. How English Works owns those mechanisms. This article uses them to make translation decisions.

The governing principle

Do not translate the smallest visible piece. Translate the smallest meaning-bearing unit that contains enough evidence to make the decision safely.

Then review the result at the next larger scale.

This two-direction movement — narrow enough to decide, wide enough to verify — is the core skill.

Where this architecture node leads next

The Translation Unit node naturally routes into later specialist branches on lexical choice, grammar and syntax, reference and cohesion, tone and register, idioms and figurative language, terminology and termbases, translation memory, AI and machine translation, post-editing, revision, quality assurance, localisation, transcreation, interpreting, literary translation, technical translation, legal translation, medical translation, academic translation, business translation, website and app localisation, SEO translation, subtitle translation and specific language pairs.

Its place in the architecture is foundational. Source analysis tells us what the source is doing. Equivalence tells us that languages do not map one-to-one. The Context Stack tells us which surrounding layers constrain meaning. The Translation Unit tells us how much of that source must be held together for one safe transfer decision.

Continue through the eduKateSG translation architecture

Master translation root: Master Art of Translation | The Complete System for Moving Meaning Between Languages

Source analysis: Master Art of Translation | How to Read a Source Text Before You Translate It

Equivalence: Master Art of Translation | How Equivalence Works When Languages Do Not Match One-to-One

Context: Master Art of Translation | The Context Stack

Vocabulary support: Vocabulary Learning Hub

English mechanism support: How English Works

Professional reference points

Current professional terminology and workflow references used for this article include Phrase documentation on translation memories and segmentation, translation-unit metadata and segmented CAT-tool workflows; memoQ documentation on joining and splitting segments; the MQM error typology; ISO 5060:2024 for translation-output evaluation; and ISO 17100:2015 for translation-service requirements. As of 2026, ISO lists a new edition of ISO 17100 as under development while the 2015 edition remains the published standard.

Final checklist: choose the unit before you choose the words

Before translating, ask: What is the visible segment? What is the actual decision? Which dependencies control it? What is the smallest span containing those dependencies? What broader context must be consulted? What production unit will be manageable? What should be stored in memory, terminology or a decision log? What larger unit must be reviewed afterward?

If the source is a phrase, do not assume the word is the unit. If the source is a sentence, do not assume the sentence is enough. If the tool shows one row, do not assume the row contains the whole problem. If the AI model can read an entire document, do not assume the entire document should be one output block. Translation quality depends on choosing boundaries deliberately.

The master skill is not translating more text at once. It is knowing what must move together.